mcp_toolset cache_control: cache the Sume tool list on Haiku 5.5

Yes, you can put a cache breakpoint on the Sume toolset. Anthropic's connector page lists cache_control as a field on mcp_toolset, described as the prompt-caching breakpoint configuration for that toolset. Haiku 5.5's overview gives the minimum cacheable prompt as 512 tokens, so a toolset that lists Sume's tools will normally clear it.

Prices from Anthropic

Cache prices are for prompts up to 100,000 tokens; above that the page lists a higher tier.

Where it goes

Add the field to the toolset object, after the config that decides which tools are present. Changing the tool list, a tool's enabled state, or anything before the breakpoint is a change to the cached prefix, so decide your allowlist first and keep it fixed during a session.

If you pin the tool list with the mcp-client-2026-09-15 beta the list is stable for the conversation by design. Sume's side of a cache hit is not affected; paid calls still need an idempotency_key, described on the tools and gates page.

What to measure

Do not guess the benefit. Send the same first turn twice and compare the usage block of the two responses: the second should show cache reads where the first showed a cache write. If it does not, something before the breakpoint changed, such as the tool list, the system prompt or the model.

Also watch the 100,000-token step in the price table. The cache read price rises from $0.01 to $0.05 per million tokens above it, so a long conversation can cross into the higher tier. The model page does not say that the tier is per request; it is described for prompts, so check the pricing page before you budget.

• Compare usage across two identical turns.
• Keep the tool list fixed during a session.
• Re-check the 512-token minimum if you trim the allowlist.